Eurographics Conference on Visualization (EuroVis) 2017 Short PaperB. Kozlíková, T. Schreck, and T. Wischgoll (Editors)

TExVis: An Interactive Visual Tool to Explore Twitter Data

Shah Rukh Humayoun1, Saman Ardalan1, Ragaad AlTarawneh2, and Achim Ebert1

1Computer Graphics and HCI Group, University of Kaiserslautern, Germany2Informaiton Technology Department, Mutah University, Jordan

AbstractExploring tweets enables us understanding people’s reaction and feedback regarding any particular event or product. Manytools have been developed to visualize Twitter data based on some criteria, e.g., keyword frequency or evolution of topics.Visualizing the relations between the keywords of the underlying Twitter data opens another window to analyze the people’sreaction towards a particular event/product. Targeting this concern, our developed tool, called TExVis (Tweets Explorer andVisualizer), visualizes important keywords (e.g., hashtags, nouns, verbs) from the underlying tweets based on their frequencyand shows the relations between them based on some criteria (e.g., the common tweets), using an extended Chord diagram.TExVis also visualizes the sentimental polarity for a better understanding of the keywords associated tweets. Further, theprovided interaction, multi-level navigation, and filtering options help the users in better exploration of the underlying tweets.A user study with 16 participants shows a high acceptance towards the tool and our approach in general.

Categories and Subject Descriptors (according to ACM CCS): H.5 [Information interfaces and presentation]: - [-]: —-H.5.2[UserInterface]: Graphical user interfaces, interaction styles

1. Introduction

Current social media platforms (e.g., Twitter, Facebook, etc.) areimportant mediums for people to express their feelings or to pro-vide their feedback towards some currently happening or recentlyhappened events. The exploration and the analysis of these feelingsand feedbacks help us understanding the overall behavior or thereaction of the community about the underlying event. Nowadays,even companies are interested in analyzing their customers’ feed-back obtained from social media to better understand trends andthe attitude towards their products. The 140-characters limitationon Twitter enforces users to write their tweets in a shortened andmore precise way, which could be more useful for the explorationand the analysis compared to other social media platforms.

Many visualization tools have been proposed in the past to ex-plore Twitter data, e.g., Nokia Internet Pulse [KLJ∗12], Spark-Clouds [LRKC10], TopicFlow [MSH∗13], Conference Monitor(CM) [SRBS12], TweetViz [SDM14], etc. Most of these tools vi-sualize Twitter data based on either high number of keywords fre-quency (e.g., [KLJ∗12]), a timeline of keywords (e.g., [SRBS12]),or evolution of topics (e.g., [MSH∗13] or [CLT∗11]). However,the relations between keywords based on some criteria (e.g., a co-occurrence relation that occurs between any two keywords if bothare in the same tweet) opens another window to explore and analyzeTwitter data. Visualizing such relations also helps users exploringand understanding people’s feelings and feedback towards the re-lated event(s) of the underlying tweets. Further, such visualizationsupport can be useful for many application domains, e.g., exploring

users’ feedback towards a product for marketing purpose or analyz-ing users’ desired functionalities from the underlying tweets.

Targeting this concern, we developed a visual tool called TExVis(Tweets Explorer and Visualizer). It visualizes not only the key-words from the underlying tweets based on their frequency but alsoshows the relations, based on the selected criteria between thesekeywords (e.g., if two keywords occur in the same tweet then theyhave a direct relation). TExVis provides the resulting visualizationthrough an extended Chord diagram that limits the extra chordscluttering, which might appear due to multiple relations’ associ-ating to the underlying nodes (i.e., keywords). Further, it visual-izes the sentimental polarities of the keywords associated tweets,which helps understanding the associated tweets’ subject. The pro-vided multi-level navigation facility, the intuitive interaction andthe filtering options help users in better exploring of the underlyingtweets. We conducted a user study with 16 participants to see howthey analyze people’s feedback towards the “Brexit” event usingthe extracted tweets of the first ten days of July 2016 from Twitter.The participants showed high interest in the exploratory tasks andprovided positive feedback towards the provided visual approach.

2. Related Work

One of the earlier work in visualizing text was done by Havre etal. [HHN00] in their famous ThemeRiver tool that visualizes thethemes’ variations over time for a collection of documents. Later,Don et al. [DZG∗07] developed FeatureLens tool that visualizes

c© 2017 The Author(s)Eurographics Proceedings c© 2017 The Eurographics Association.

DOI: 10.2312/eurovisshort.20171149

Humayoun et al. / TExVis: An Interactive Visual Tool to Explore Twitter Data

text collection at several levels of granularity to enable users explor-ing interesting text patterns based on length and frequency, whileCui et al. [CLT∗11] in their TextFlow work used a semi-supervisedclustering technique for the topic creation and represented the topicconvergence and divergence using the river metaphor. As socialmedia platforms (e.g., Twitter) provide large volume of real-timedata; therefore, many researchers have focused on finding severaltechniques to visualize Twitter stream and hashtags. Few exam-ples are: Kaye et al. [KLJ∗12] developed a tool to visualize theevolution of Twitter discussions with a time series of stacked tagclouds. TopicFlow [MSH∗13] [NMM11] visualizes the evolutionof tweets to help understanding statistical topic modeling. Sopan etal. [SRBS12] provided an analysis of academic conferences hash-tags over time to analyze conferences’ trends. SparkClouds tool,developed by Lee et al. [LRKC10], integrates spark lines into thecloud tags to convey the trends between the multiple tag clouds.Stojanovski et al. [SDM14] developed the TweetViz tool that repre-sents topic distribution in a set of tweets to allow users searching forany hashtag or keywords in the proposed interface. Thom et al. usedScatterBlogs to visualize geo-located Twitter messages [TBK∗12]and to study crisis intelligence [TKE∗15], e.g., using sentiment vol-umes of geo-located tweets. While Dork et al. [DGWC10] visual-ized Twitter data in three modes: topics over time through TopicsStreams layout, people and their activity through spiral layout, andpopularity of event photos through Image Cloud. Most of the pre-vious work focused on keywords frequency or evolution of topics.None of them investigated the impact of different relations betweenthe keywords (based on some relation criteria). Visualizing theserelations could help the users to explore and understand people’sfeelings and feedback towards a particular event, product, or term.

3. The Enhanced Chord Diagram

The Chord diagram is a radial layout, initially popularized by TheNew York Times to show the relations between Genomes usingthe Circos package [KSB∗09] [TNYT]. Radial (circular) layoutsproduce compact visualizations and use space efficiently, as theysupport a larger data domain on a squared area than rectangularor square layouts provide [KSB∗09] [Krz]. They encourage theeye movement to proceed along the curved lines rather than a zig-zag fashion in a square or rectangular figure, which helps view-ers to better understand and explore the underlying data [Krz].Also, they can show the flow on relations between pairs more intu-itively [KSB∗09]Due to these reasons, we selected Chord diagramrather than any rectangular/square layout (e.g., matrix layout).

In standard Chord diagrams, data elements (also called nodes orarcs) are arranged in a circular fashion and relations (also calledchords) are drawn between the nodes. Mostly, chords associated toa node are mutually exclusive due to their association with differentdata in the underlying data-element; hence, no chords overlappinghappens at the node side (see Fig. 1.a). However, sometimes chordsassociated to a node may not be mutually exclusive, which createschords overlapping at the associated node side. This can cause acluttering issue in the resulted visualization. Handling clutteringresulted from many-to-many relations in the visualization is a chal-lenging task, which has been targeted by some researchers in thepast for different visualization techniques (e.g., for matrix-base vi-

sualization [YDGM17] or for correlation map [ZMZM15]). How-ever, as per our knowledge no one handled in the past the clutter-ing issue in Chord diagram resulting from non-mutually exclusivemany-to-many relations.

Figure 1: (a) A standard Chord diagram taken from [Bos], (b) re-lations associated to nodes in TExVis extended Chord diagram.

To deal with the cluttering issue in our case, we propose an ex-tension to the standard Chord diagram (see Fig. 1.b), which weuse in our TExVis tool. In our case, the width of a node repre-sents the weight value of the data element (e.g., frequency of a key-word), while the height of a node represents the number of associ-ated chords or relations (e.g., co-occurrence relations between thiskeyword and other keywords in the underlying tweets). The widthof a chord represents the weight value of this relation (e.g., fre-quency of co-occurrence relation between two keywords). In orderto avoid extra chords cluttering, we order the chords based on theirweights, such that the chord with highest weight value is placed atthe bottom, while the second next one is placed above the previousone, and so on till the chord with the least weight value. However,the chord with the highest weight value starts a little below theupper/outer side of the specified node, the second one starts fewpixels downwards, and so on (see Fig. 1.b). In this case, the nodewith the least weight value starts just above the inner boundary ofthe node. In this proposed solution, no chord is hidden behind theother chords; therefore, it provides more readable visualization.

4. TExVis: Tweets Explorer and Visualizer

Our developed TExVis tool visualizes the frequent keywords inthe underlying tweets, the relations between these keywords us-ing some particular criteria, and the sentimental polarities of theassociated tweets. The web-based client side was developed usingHTML, CSS, and JavaScript to provide the visual view as well asthe interaction and filtering options, while the server side was de-veloped in C#.Net to manage and process the data.

TExVis fetches the tweets from Twitter using the Tweetinvi andTwitter REST APIs. It fetches the data per user’s request using arequesting loop. Then it separates interesting words (we call themkeywords or tokens) based on hashtags, nouns and verbs using theApache OpenNLP, which is a Natural Language Processing (NLP)library. If a non-noun or a non-verb hashtag is used frequently inthe previous retrieved tweets then it is marked as a global hash-tag, called g-hashtag, and the corresponding tokens are also sep-arated in the current extracted Twitter data. Sentiment classifier isa term used to classify the text based on the contained sentimentalpolarities (e.g., positive, negative, or neutral) [LLC∗10] [PLV02],

c© 2017 The Author(s)Eurographics Proceedings c© 2017 The Eurographics Association.

152

Humayoun et al. / TExVis: An Interactive Visual Tool to Explore Twitter Data

Mouse hovering over a node

Clicking on a tweet

Mouse hovering over a tweet

Navigating “vote” and then “week”

Relation based on words-similarity

(a)

(b)

(c)

(d)

(e)

brexit vote week

DATASET:

RelationType:

Colorization:

RelationValue(%):

Keywords:

DateFrom:

DateTo:

WordType:

From:

brexit

Overall Value:

CurrentValue: Fri., 01.07.2016, 00:00:00 --- Fri., 10.07..2016, 23:59:59

Fri., 01.07.2016, 00:00:00 --- Fri., 10.07..2016, 23:59:59

Tweet (687) by “User3988”

NumberofNodes:

Figure 2: (a) TExVis tool with the left filter panel, the central visualization panel, and the right tweet panel, (b) the Chord diagram based onthe words similarity relations, (c) navigating the keyword “vote” and then keyword “week”, (d) showing only the relations of the keyword“uk”, and (e) showing only the associated nodes and relations of a selected tweet.

which is useful for getting an overall opinion towards the text sub-ject [PLV02]. TExVis uses Aylien.TextApi [ATAA] for each re-trieved tweet, which returns the sentimental polarity value of thetweet, along with the polarity confidence value (between 0% to100%) to show the confidence level of the stated polarity value.

Figure 2.a shows the overall view of TExVis client-side. It con-sists of three parts: the filter panel at the left side, the visualizationpanel in the center containing the extended Chord diagram built us-ing the d3.js library, and the tweet panel at the right side. For a proofof concept, here we use the tweets extracted from Twitter using thekeyword “brexit”, that was a popular hashtag in July 2016 referringto the UK referendum about its quitting from EU. Our extracteddata consists of 41,199 tweets (with 56,701 distinct keywords in allcategories) between July 01 to July 10, 2016. We assigned randomID numbers in order to make users’ IDs anonymous.

In order to explore and analyze the underlying Twitter data invisual form, the central part of the visualization panel is dedicatedto our extended Chord diagram. In this diagram, the node’s widthrepresents the keyword’s frequency while the node’s height is basedon the number of associated relations to this node. The relations be-tween the nodes are decided through the given criteria. Currently,TExVis supports relations based on two criteria: the co-occurrencecriteria in which a relation between two nodes (keywords) occursif both belong to the same tweet, and the chord (relation) widthdepends on the frequency of this co-occurrences in the underlyingtweets (see Fig. 2.a); and the words similarity criteria in which thechord (relation) width indicates the value of the words similarity(range between 0 to 1) of the connected nodes (see Fig. 2.b) thatis calculated in TExVis using the WordNet.Net [SC] library, whichacts as a semantic dictionary for English lexical tokens. The colorsof chords associated to a node ranges from darker to lighter fromthe wider to the thinner chords respectively with the same nodecolor. Also, the chord color between two nodes depends on thelarger associated node (e.g., a chord between “uk” and “vote” inFig. 2.a has the same color as of the “uk” node). However, mouse

hovering over a particular node fades all other nodes’ relations andchanges the colors of this node’s relations according to the oppositeassociated nodes (see Fig. 2.d). Mouse hovering over a particularnode or a relation also brings a tooltip to show further details (e.g.,no. of associated tweets, percentage, etc.). Further, we provide arcsouter-side of nodes to show the sentimental polarity (here, greenrepresents positive, blue represents neutral and red represents neg-ative) of all the associated tweets to each node. The color opacityshows the average confidence level for each polarity, i.e., 100% isthe darkest color it gets. The length of each color in the arc repre-sents the frequency of this polarity in the associated tweets.

Users can navigate the Chord diagram on-demand by selectingthe navigation option from the menu bar (appears by clicking on aparticular node), which results a new Chord diagram as a next levelof details. For example, navigating “vote” and then “week” visual-izes the data in the resulting Chord diagram related to only thosetweets that have “brexit”, “vote”, and “week” together in them (seeFig. 2.c). This helps users exploring the tweets based on intersec-tion of keywords. A navigation path is also shown at the top, whichis used for going back to any previous level of details. TExVisalso provides the navigation option through a chord; however, inthis case user goes two levels down, e.g., Fig. 2.c navigation canbe achieved by selecting the option from the chord between the“vote” and “week” nodes.

The right-side tweet panel shows tweets associated to the cur-rent Chord diagram. User can filter it to see tweets only relatedto a particular node or chord. Clicking a particular tweet opens adown space to show full text and all the associated keywords, whilemouse hovering over a particular tweet fades all non-associatednodes and relations in the central Chord diagram (see Fig. 2.e).

TExVis provides a number of filtering options in the left side fil-ter panel, for example: selecting the main extracted dataset basedon the extracted keyword (e.g., “brexit”), selecting the relation type(e.g., co-occurrence or words similarity), navigating the visual-ization based on a given keyword, visualizing the Chord diagram

c© 2017 The Author(s)Eurographics Proceedings c© 2017 The Eurographics Association.

153

Humayoun et al. / TExVis: An Interactive Visual Tool to Explore Twitter Data

with the selected number of the most frequent nodes (e.g., Fig. 2shows the Chord diagram with 20 high frequent keywords), filter-ing through date and time, and filtering based on the keyword-type(e.g., nouns, verbs, or hashtags). A time-bar is also provided at thebottom side of the central Chord diagram to filter the current viewbased on time interval.

5. The User Study

We conducted a user study with 16 participants (6 females), aged24-33 (M = 28.9). The used dataset was the earlier described“brexit” dataset. The study goal was to check how users ana-lyze and understand people feedback regarding the “brexit” eventthrough exploring the provided visualization. We were also inter-ested to see how different relationship types between the keywordsinfluence on users in understanding and exploring the underlyingtweets. Finally, we wanted to know users’ reaction towards our ap-proach and the visualization. Based on this, we defined the studywith a total of six tasks where the purpose of the first four tasks (i.e.,getting information, navigating the visualization, data filtering, andunderstanding keywords’ relations) was to make the participantswell aware of the TExVis visualization and the provided interac-tion and filtering options. The last two tasks were exploratory innature where participants were asked to analyze people’s feedbackand reaction towards the “brexit” event, first from the perspective ofnavigating to “eu” and then from the perspective of further navigat-ing to “vote”. The study was done between-subjects manner, whereeight of the randomly selected participants performed the experi-ment using the co-occurrence relation type while the other eightperformed it using the words similarity relation type. At the end ofthe experiment, participants were asked to answer a 6 closed-endedquestionnaire (using Likert-scale from 1 to 5) and their feedback ingeneral. Each experiment lasted no more than 1 hour.

In the first four tasks, both groups performed approximately thesame in terms of accuracy, i.e., 90.63%, 98.94%, 100%, and 100%by the co-occurrence group compared to 95.75%, 100%, 100%, and100% by the words similarity group. However, the co-occurrencegroup’s performance time was overall better in these tasks, i.e.,208.75, 171.88, 214.38, and 98.75 seconds compared to 278.75,185.63, 246.25, and 88.13 seconds by the words similarity group. Inthe last two exploratory-natured tasks, participants in both groupsshowed high interests and provided interesting feedback. Few ex-amples of feedback from the co-occurrence group are: “People arestill in shock of the referendum event and do not have clear ideaabout the consequences”, “There is interest in signing a petitionto protest the results”, and “Many people are linking the situationback to WW2”. Two examples of the words similarity group feed-back are: “After navigating to vote, people are talking about ‘hate’,‘fear’ and ‘crime’” and “It is interesting to see ‘nato’, ‘crimes’,‘hate’, ‘police’, ‘fear’, ‘Youtube’, and ‘Brussels’. Something re-lated to terrorism had happened at the same time? ”

We found out that the co-occurrence group highly relied on theco-occurrence relation for the analysis, especially where they wererelating two events. Most of the participants in this group had asimilar approach, i.e., they found the most occurring keyword pairsand then tried to find out a reason behind it. Few of them usednavigation option to go further to their own decided next level of

details in order to focus on some specific topics. From the feed-back, it is clear that they used features like sentiment polarity, co-occurrence relation, words frequency, and navigation to understandpeople’s behavior and to give the answer. Overall, they were ableto find out various aspects of the “brexit” event. Some were evenshocked to see the sentiment polarity of tweets related to partic-ular keywords. Finding the relation of “brexit” with some otherevents (e.g., “WW2”) helped them understanding people’s reactionand feedback. While the second group had some other perspectives.Although they used the sentiment polarity, the navigation, and thewords frequency; however, it seems that they hardly made any con-clusion out of the similarity relation. It is because the words simi-larity relation shows that how much two words are semantically orlexically similar and this does not help the users to relate it to anevent. However, the similarity relation can be useful in some otherscenarios, e.g., clarifying two confusing words and knowing whichones in a certain context could be used by people (e.g., the case of‘Geek’ vs. ‘Nerd’ by Settles [Set13]).

In closed-ended questions, most of participants from both groupseither agreed or strongly agreed with the statements. In the case ofintuitiveness of the visualization, 7 agreed and 8 strongly agreed. Inthe case of clearness and understandability, 7 agreed and 6 stronglyagreed. When asked about visualization support for the analysispurpose, 5 agreed and 9 strongly agreed with it. However, 5 agreedand 1 strongly agreed in the co-occurrence group towards the us-age of co-occurrence relation compared to 3 agreed and 2 stronglyagreed the second group towards the usage of words similarity re-lation. There is some disagree of usage of word similarity rela-tion that we can understand from the given scenario. This and theexploratory-natured tasks’ feedbacks indicate that different relationtypes suit different exploration scenarios, e.g., co-occurrence rela-tion type suited more to our experiment scenario compared to thewords similarity relation type. Therefore, it is recommended to firstinvestigate whether a particular relation type could help in explor-ing and analyzing the underlying scenario. Most of the participants(5 agreed and 8 strongly agreed) favored the semantics polarity op-tion as well as high positive feedback towards using the tool in fu-ture (4 agreed and 10 strongly agreed). In the open-ended feedback,few participants suggested to show initially only the important re-lations and then the remaining ones on demand. However, all ofthem provided high positive feedback about the tool, the visualiza-tion, and the approach. Few also asked to make the tool public.

6. Conclusion and Future Work

The presented TExVis tool enables the visual exploration of Twit-ter data through keywords frequency, keywords relation, and asso-ciated tweets’ sentimental polarities using the proposed extendedChord diagram. The conducted user study indicates that the usedkeywords’ relation type is useful when it supports the analysis ofthe underlying events or scenarios. In the future, we aim to provideadditional facilities in TExVis, such as: selecting the relation basedon other criteria, support of other social media platforms, showingonly the important relations initially and further other relations onlyon demand, and visual comparison of people reaction about two ormore events. We also plan to conduct detailed user studies on largerscale to generalize our findings.

c© 2017 The Author(s)Eurographics Proceedings c© 2017 The Eurographics Association.

154

Humayoun et al. / TExVis: An Interactive Visual Tool to Explore Twitter Data

References

[ATAA] AYLIEN-TEXT-ANALYSIS-API:. URL: http://docs.aylien.com/docs/introduction. 3

[Bos] BOSTOCK M.: Chord diagram. URL: https://bl.ocks.org/mbostock/4062006. 2

[CLT∗11] CUI W., LIU S., TAN L., SHI C., SONG Y., GAO Z., QU H.,TONG X.: Textflow: Towards better understanding of evolving topics intext. IEEE Transactions on Visualization and Computer Graphics 17, 12(Dec. 2011), 2412–2421. URL: http://dx.doi.org/10.1109/TVCG.2011.239, doi:10.1109/TVCG.2011.239. 1, 2

[DGWC10] DORK M., GRUEN D., WILLIAMSON C., CARPENDALES.: A visual backchannel for large-scale events. IEEE Transactionson Visualization and Computer Graphics 16, 6 (Nov. 2010), 1129–1138. URL: http://dx.doi.org/10.1109/TVCG.2010.129,doi:10.1109/TVCG.2010.129. 2

[DZG∗07] DON A., ZHELEVA E., GREGORY M., TARKAN S., AUVILL., CLEMENT T., SHNEIDERMAN B., PLAISANT C.: Discovering inter-esting usage patterns in text collections: Integrating text mining with vi-sualization. In Proceedings of the Sixteenth ACM Conference on Confer-ence on Information and Knowledge Management (New York, NY, USA,2007), CIKM ’07, ACM, pp. 213–222. URL: http://doi.acm.org/10.1145/1321440.1321473, doi:10.1145/1321440.1321473. 1

[HHN00] HAVRE S., HETZLER B., NOWELL L.: Themeriver: visual-izing theme changes over time. In IEEE Symposium on InformationVisualization 2000. INFOVIS 2000. Proceedings (2000), pp. 115–123.doi:10.1109/INFVIS.2000.885098. 1

[KLJ∗12] KAYE J. J., LILLIE A., JAGDISH D., WALKUP J., PARADAR., MORI K.: Nokia internet pulse: A long term deployment and itera-tion of a twitter visualization. In CHI ’12 Extended Abstracts on HumanFactors in Computing Systems (New York, NY, USA, 2012), CHI EA’12, ACM, pp. 829–844. URL: http://doi.acm.org/10.1145/2212776.2212856, doi:10.1145/2212776.2212856. 1, 2

[Krz] KRZYWINSKI M. I.: Benefits of a circular layout. circos. URL:http://circos.ca/intro/circular_approach/. 2

[KSB∗09] KRZYWINSKI M., SCHEIN J., BIROL I., CONNORS J., GAS-COYNE R., HORSMAN D., JONES S. J., MARRA M. A.: Circos: aninformation aesthetic for comparative genomics. Genome research 19, 9(2009), 1639–1645. 2

[LLC∗10] LI S., LEE S. Y. M., CHEN Y., HUANG C.-R., ZHOU G.:Sentiment classification and polarity shifting. In Proceedings of the 23rdInternational Conference on Computational Linguistics (Stroudsburg,PA, USA, 2010), COLING ’10, Association for Computational Linguis-tics, pp. 635–643. URL: http://dl.acm.org/citation.cfm?id=1873781.1873853. 2

[LRKC10] LEE B., RICHE N. H., KARLSON A. K., CARPENDALE S.:Sparkclouds: Visualizing trends in tag clouds. IEEE Transactions onVisualization and Computer Graphics 16, 6 (Nov. 2010), 1182–1189.URL: http://dx.doi.org/10.1109/TVCG.2010.194, doi:10.1109/TVCG.2010.194. 1, 2

[MSH∗13] MALIK S., SMITH A., HAWES T., PAPADATOS P., LIJ., DUNNE C., SHNEIDERMAN B.: Topicflow: Visualizing topicalignment of twitter data over time. In Proceedings of the 2013IEEE/ACM International Conference on Advances in Social NetworksAnalysis and Mining (New York, NY, USA, 2013), ASONAM ’13,ACM, pp. 720–726. URL: http://doi.acm.org/10.1145/2492517.2492639, doi:10.1145/2492517.2492639. 1, 2

[NMM11] NALLAPATI R., MCFARLAND D. A., MANNING C. D.: Top-icflow model: Unsupervised learning of topic-specific influences of hy-perlinked documents. In AISTATS (2011), pp. 543–551. 2

[PLV02] PANG B., LEE L., VAITHYANATHAN S.: Thumbs up?: Sen-timent classification using machine learning techniques. In Pro-ceedings of the ACL-02 Conference on Empirical Methods in Natu-ral Language Processing - Volume 10 (Stroudsburg, PA, USA, 2002),

EMNLP ’02, Association for Computational Linguistics, pp. 79–86.URL: https://doi.org/10.3115/1118693.1118704, doi:10.3115/1118693.1118704. 2, 3

[SC] SIMPSON T., CROWE M.: Wordnet.net. URL: http://www.ebswift.com/wordnetnet.html. 3

[SDM14] STOJANOVSKI D., DIMITROVSKI I., MADJAROV G.: Tweet-viz: Twitter data visualization. Proceedings of the Data Mining and DataWarehouses (2014). 1, 2

[Set13] SETTLES B.: On “geek" versus “nerd", 2013. URL:https://slackprop.wordpress.com/2013/06/03/on-geek-versus-nerd/. 4

[SRBS12] SOPAN A., REY P. J., BUTLER B., SHNEIDERMANB.: Monitoring academic conferences: Real-time visualiza-tion and retrospective analysis of backchannel conversations. InProceedings of the 2012 International Conference on Social In-formatics (Washington, DC, USA, 2012), SOCIALINFORMAT-ICS ’12, IEEE Computer Society, pp. 62–69. URL: http://dx.doi.org/10.1109/SocialInformatics.2012.20,doi:10.1109/SocialInformatics.2012.20. 1, 2

[TBK∗12] THOM D., BOSCH H., KOCH S., WORNER M., ERTL T.:Spatiotemporal anomaly detection through visual analysis of geolo-cated twitter messages. In Proceedings of the 2012 IEEE Pacific Vi-sualization Symposium (Washington, DC, USA, 2012), PACIFICVIS’12, IEEE Computer Society, pp. 41–48. URL: http://dx.doi.org/10.1109/PacificVis.2012.6183572, doi:10.1109/PacificVis.2012.6183572. 2

[TKE∗15] THOM D., KRÃIJGER R., ERTL T., BECHSTEDT U., PLATZA., ZISGEN J., VOLLAND B.: Can twitter really save your life? acase study of visual social media analytics for situation awareness. In2015 IEEE Pacific Visualization Symposium (PacificVis) (April 2015),pp. 183–190. doi:10.1109/PACIFICVIS.2015.7156376. 2

[TNYT] THE-NEW-YORK-TIMES: Close-ups of the genome, species byspecies by species. IEEE Transactions on Visualization and ComputerGraphics. URL: http://www.nytimes.com/imagepages/2007/01/22/science/20070123_SCI_ILLO.html. 2

[YDGM17] YANG Y., DWYER T., GOODWIN S., MARRIOTT K.: Many-to-many geographically-embedded flow visualisation: An evaluation.IEEE Transactions on Visualization and Computer Graphics 23, 1 (Jan2017), 411–420. doi:10.1109/TVCG.2016.2598885. 2

[ZMZM15] ZHANG Z., MCDONNELL K. T., ZADOK E., MUELLERK.: Visual correlation analysis of numerical and categorical data onthe correlation map. IEEE Transactions on Visualization and ComputerGraphics 21, 2 (Feb 2015), 289–303. doi:10.1109/TVCG.2014.2350494. 2

c© 2017 The Author(s)Eurographics Proceedings c© 2017 The Eurographics Association.

155